Efficient material development using a combination of thin-film growth techniques and machine-learning approaches in energy applications
Efficient material development using a combination of thin-film growth techniques and machine-learning approaches in energy applications
复制标题
在能源应用中结合薄膜生长技术和机器学习方法进行高效材料开发
DOI：
10.11470/jsaprev.220401
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
後藤真宏
中科院分区：
文献类型：
--
作者：
大久保勇男;後藤真宏
2011年，美国启动了材料基因组计划[1]。这一研究趋势已在全球范围内蔓延，即使在今天仍然是材料研究的主要研究课题。最初，主要研究课题包括构建包含第一原理计算结果的大型数据库，开发用于材料研究的机器学习代码，以及结合第一原理计算和机器学习的研究。然而，今天，这些方法已经达到了传播到各种形式的实验研究的阶段，并有积极的努力，涉及实验研究人员。正在尝试将机器学习引入各种形式的材料研究中，不仅用于数据驱动的研究，涉及选择和筛选可预期具有所需特性的候选材料，而且还用于分析/优化工艺参数和分析/数据库形成各种测量结果。将机器学习引入到关于薄膜制造过程的实验研究中也不例外，我们相信这一过程的自然进展包括引入机器学习，以有效地获得表现出所需特性的薄膜样品。薄膜制造工艺，其中可以从具有原子层厚度的超薄膜制造具有微米量级厚度的薄膜样品，是材料研究的基本方法，并且不仅用于学术基础研究的研究领域，而且用于各种工业领域。气相薄膜制造工艺具有优点，例如不同组成的薄膜的平行合成和由于引入高通量方法而通过组成梯度的集成。然而，表现出所需性质的薄膜样品的制造涉及需要优化多种类型的薄膜生长参数和相对高的操作成本的问题，这涉及设备维护和管理。减少薄膜制造过程的负载并提高其效率不仅是合成和寻找材料的方法，而且是器件制造所必需的过程，都是一个紧迫的问题。本文介绍了两个将机器学习引入到薄膜制造过程中的材料开发研究中的研究实例，并探讨了作者所在研究小组将机器学习方法应用于热电薄膜制造的情况。
The last year marks 10 years since the start of the Materials Genome Initiative [1] in the United States in 2011. This research trend has spread globally and has remained a major research topic in materials research even today. Initially, the major research topics comprised the construction of largescale databases that contain first-principles calculation results, development of machine-learning code for materials research, and research that combines first-principles calculations and machine learning. However, today, these methods have reached the stage of dissemination to various forms of experimental research, and there are active efforts that involve experimental researchers. Attempts are being made to introduce machine learning into various forms of materials research, not only for the data-driven research that involves the selection and screening of candidate materials that can be expected to have the desired properties, but also the analysis/optimization of process parameters and the analysis/database formation of various measurement results. The introduction of machine learning into experimental research regarding the thin-film fabrication process is no exception, and we believe the natural progression of this comprises the introduction of machine learning in order to efficiently obtain thin-film samples that exhibit the desired properties. The thin-film fabrication process, in which a thin-film sample can be fabricated with a thickness on the order of micrometers from an ultra-thin film with an atomic-layer thickness, is an essential method for materials research and is used not only in the research field of academic basic research but also in various industrial fields. The vapor-phase thin-film fabrication process has advantages such as parallel synthesis of thin films of different compositions and integration by composition gradient owing to the introduction of the highthroughput method. However, the fabrication of a thin-film sample that exhibits the desired properties involves the issues of the need to optimize multiple types of thin-film growth parameters and the relatively high operating cost, which involves equipment maintenance and management. Reducing the load and improving the efficiency of the thin-film fabrication process is an urgent issue not only as a method for synthesizing and searching for materials but also as a process that is essential for device fabrication. In this paper, we introduce two research examples that introduced machine learning into material development research via the thin-film fabrication process, and we investigate the application of this method to thermoelectric thin-film fabrication, which was performed by the authors’ research group.